In this paper, we report experiments on the use of distributed binary representations in an architecture based on several TDNN-DTW hybrid systems in parallel, for word recognition. Each TDNN is trained on a different set of binary distinctive features chosen in an arbitrary manner and yield different confusion matrices on the training set. These confusion matrices are used to obtain a global recognition decision by combining all system outputs. We investigate two possibilities which we call multiplicative mode and fusion mode. We show that the proposed architecture is able to provide better word recognition rate than a simple TDNN-DTW hybrid system trained on the 1-out-of-N representation and that it is comparable with a system trained on binary phonetic features. The fusion mode turns out to be an efficient strategy for handling word recognition in our parallel architecture.
Bibliographic reference. Puzrla, Premysl / Bimbot, Frédéric / Windheuser, Christoph (1995): "Distributed binary representations for word recognition by TDNN-DTW hybrid systems", In EUROSPEECH-1995, 2175-2178.